From: Greg Banks Subject: Re: Performance udp vs. tcp Date: Mon, 23 Aug 2004 20:36:37 +1000 Sender: nfs-admin@lists.sourceforge.net Message-ID: <20040823103637.GK10775@sgi.com> References: <4129BF94.20902@bio.ifi.lmu.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: nfs@lists.sourceforge.net Return-path: Received: from sc8-sf-mx1-b.sourceforge.net ([10.3.1.11] helo=sc8-sf-mx1.sourceforge.net) by sc8-sf-list2.sourceforge.net with esmtp (Exim 4.30) id 1BzCC1-00010t-Kt for nfs@lists.sourceforge.net; Mon, 23 Aug 2004 03:36:49 -0700 Received: from omx3-ext.sgi.com ([192.48.171.20] helo=omx3.sgi.com) by sc8-sf-mx1.sourceforge.net with esmtp (Exim 4.34) id 1BzCC0-0002gA-5S for nfs@lists.sourceforge.net; Mon, 23 Aug 2004 03:36:49 -0700 To: Frank Steiner In-Reply-To: <4129BF94.20902@bio.ifi.lmu.de> Errors-To: nfs-admin@lists.sourceforge.net List-Unsubscribe: , List-Id: Discussion of NFS under Linux development, interoperability, and testing. List-Post: List-Help: List-Subscribe: , List-Archive: On Mon, Aug 23, 2004 at 11:57:40AM +0200, Frank Steiner wrote: > Hi, > > I just ran some bonnie++ tests to measure the nfs performance of udp vs. > tcp, expecting tcp to be better (after reading this list and some > newsgroups). > Amazingly, it turned out that udp with rsize/wsize 16k was the best, taking > the average over all tests. tcp was a little better in some of the > large-file > tests, but especially with creating/stating/deleting 16 thousands of small > file, sequentially and randomly, udp was better than tcp in every test > (between 5-15 % faster). I rann all the tests also with differenct tcp > windows > sizes, and a larger window size speeded up udp and tcp more or less in the > same way. With a metadata-heavy workload like this the [rw]size makes little difference. > I was very amazed by this, but I guess this could be caused by our network, > which is a fully switched "100mbit for client/1gbit for servers" network, > so we have almost no packet loss. > > If I got the discussions on the list and in the newsgroups right, on a > lossy network udp would slow down due to the packet loss and resends. > > So, can my result be correct for a really fast and stable network? Or is > it some "bonnie++ likes udp" effect? :-) There's nothing unusual here. On a low-latency uncongested unlossy network, it's not surprising that UDP is faster. As soon as any of those factors goes away, TCP starts winning. Greg. -- Greg Banks, R&D Software Engineer, SGI Australian Software Group. I don't speak for SGI. ------------------------------------------------------- SF.Net email is sponsored by Shop4tech.com-Lowest price on Blank Media 100pk Sonic DVD-R 4x for only $29 -100pk Sonic DVD+R for only $33 Save 50% off Retail on Ink & Toner - Free Shipping and Free Gift. http://www.shop4tech.com/z/Inkjet_Cartridges/9_108_r285 _______________________________________________ NFS maillist - NFS@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/nfs